Ajay Kumar on running MCP in production — Agents After Dark

Skip to content Agents After Dark<br>Ajay Kumar on running MCP in production<br>What happens when your customer is an AI agent, not a person?<br>AK Ajay Kumar Head of MCP Services, InfoTrack 18 July 2026 ·with Matt Doughty

YouTube Spotify Apple Podcasts<br>In this episode<br>Why InfoTrack built a dedicated MCP business unit instead of a side experiment<br>How an enterprise MCP gateway uses a registry and search-and-invoke pattern<br>Why human-in-the-loop elicitation matters once an agent can spend money<br>Why hallucination is an agent problem, since MCP itself is deterministic

Full transcript click to expand Transcript coming soon.

Take the next step<br>See where your organisation actually stands.<br>The agent readiness assessment scores your data, integration, and governance maturity in about four minutes, then tells you the highest-leverage move to make next.<br>Take the readiness assessment &rarr; Read the Agent Brief

Share this episode<br>Copied

About the guest<br>AK Ajay Kumar<br>Head of MCP Services, InfoTrack

Ajay Kumar is Head of MCP Services at InfoTrack, where he leads the company's Model Context Protocol business unit and shapes how AI moves from experimentation into core production workflows. With over 17 years across banking, prop-tech, insurance, and government, he brings an engineering-first perspective on identity, architecture, and enterprise-scale delivery.<br>Connect on LinkedIn &rarr;<br>Related reading<br>Human-in-the-Loop Patterns That Earn an Agent Autonomy<br>Building New Security Perimeters: How Agent Tool Access Changes Your Risk Model<br>When Agent Token Bills Exceed Business Value: Cost Governance Before First Deployment

From this episode<br>Frequently asked questions<br>What is the Model Context Protocol (MCP)? MCP is a standard that lets AI agents discover and call external services and data in a structured, deterministic way. It turns products that people once clicked through into services that software agents can invoke directly.

How does an enterprise MCP gateway work in practice? It centres on a registry of available tools plus a search-and-invoke pattern, where agents find the right service and call it. Vectorised tools help cut token costs across hundreds of services.

Why does human-in-the-loop elicitation matter for agents? The moment an agent can order, transact, or spend money, a human confirmation step becomes important for control and accountability. Ajay notes that most major providers still do not support elicitation out of the box.

Is hallucination an MCP problem? No. MCP itself is deterministic, so hallucination is an agent problem rather than a protocol problem. The reliability question sits with how the agent reasons, not with the tool calls.

What separates companies experimenting with agents from those scaling them? Roughly 62% of companies are experimenting while only about 28% are scaling. The episode looks at the architecture, security, and go-to-market decisions that move an organisation from pilots to production.

Stay ahead of the curve<br>Get frameworks, playbooks, and insights on agentic governance delivered to your inbox.<br>Subscribe No spam. Unsubscribe anytime. A resource by Prefactor.<br>Almost there — check your inbox to confirm your subscription.